DRC-Hubo Walking on Rough Terrains

Hongfei Wang and Yuan F. ZhengElectiral and Computer Engineering

Ohio State UniversityColumbus, OH 43210, USAEmail: wang.3185@osu.edu

Youngbum Jun and Paul OhMechanical Engineering and Mechanics

Drexel UniversityPhiladelphia, PA 19104, USA

Email: yj55@drexel.edu

Abstract— Up to now humanoid robots have been designedprimarily for walking on flat surfaces. In the future, humanoidrobots are required to replace human beings to operate innatural or damaged man-engineered environments. In the 2013DARPA Robotics Challenge, the robots are required to walkthrough several type of rough terrains. In this scenario, therobot will be challenged to keep balance and fulfill the taskswhile walking. We have developed several balance gaits andassociated controllers. The latter collaborate with a computervision system to enable our humanoid robot DRC-Hubo to walkover rough terrains. Both theoretical and experimental resultsare presented to verify the approach.

I. INTRODUCTIONIn recent years a large number of humanoid robots have

been developed in the world [1][2][3]. The original purposeis for the robot to walk on even floor but not on roughterrains. The latter however are what human beings encounteroften. It is a difficult task for humanoid robots to negotiaterough terrains. The reason is simple: maintaining stablelocomotion of humanoid robots is extremely challengingeven on flat floors as proven in reality. Let alone unevensurfaces. Today emerging applications are pushing humanoidrobots to walk in natural or man-made environments, whichdo not have surfaces prepared for humanoid robots. Wehave developed technologies to integrate balance controltechnologies and vision system into our DRC-Hubo robot,a humanoid robot originally developed by KAIST in SouthKorea. The research is inspired by the 2013 DARPA RoboticsGrand Challenge (DRC). Our ultimate goal is to enableDRC-Hubo to walk to the finishing line in DRC roughwalking event.

Surveying the existing literature, Zheng investigated bipedclimbing slopes in an early work [4] and developed ski-type gait for biped walking with improved stability for roughterrain walking [5][6]. The work dealt with the challengingissue of sudden transition of the surface from level to slope.Force sensors underneath the feet were used to detect thetransitions. More approaches for walking on uneven andinclined floor are also proposed in the recent years. Kimet al described a dynamic walking control algorithm thatimplements various online controllers to cope with local and

This work was supported in part by the Defense Advanced ResearchProjects Agency (DARPA) award #N65236-12-1-1005 for the DARPARobotics Challenge.

Fig. 1. DRC-Hubo walking up and down on ramp in DARPA RoboticChallenge

global inclinations of the floor based on an enhanced versionof a previously proposed dynamic walking algorithm [7].In a more recent work, Manchester et al proposed a newmethod for biped to walk on uneven terrain. The provably-stable feedback control strategy is based on arbitrary non-periodic trajectories arriving in real-time from an onlinemotion planner [8]. For biped to walk on generic roughterrains, approaches have been primarily on the optimal orminimization control of compass gait pattern [9][10][11].

The structure of the paper is as follows. In the nextsection we will introduce DRC-Hubo robot and the DRCrough walking event setup. In Section 3 and 4, the balancecontrollers and the vision system developed for the robot arediscussed. In Section 5 we will present experimental results.The paper will be concluded in Section 6.

II. AN INTRODUCTION OF DRC AND DRC-HUBOThe Department of Defense’s strategic plan identifies

requirements to extend aid to victims of natural or man-madedisasters and conduct evacuation operations. However, somedisasters like 2011 Fukushima accident, due to grave risksto the health and wellbeing of rescue and aid workers, provetoo great in scale or scope for timely and effective humanresponse. The DARPA Robotics Challenge (DRC) is held toadvance the current state of the art in robotics technology bycompeting in eight events [12]. Eight events are specified:Vehicle driving, Rough Terrain walking, Ladder climbing,

978-1-4799-4605-1/14/$31.00 ©2014 IEEE

Fig. 2. Hubo2 (Left): previous version. DRC-Hubo (Right): latest version

Debris cleaning, Door opening, Wall breaking, Valve turningand Hose installation. Each event has a total of four points,thus the total score is thirty-two points. For each event,the setup time is constrained to be fifteen minutes, and theoperation time is limited to be thirty minutes. If the team runsout of the setup time, the extra setup time will be countedas operation time. So the maximum operation time is thirtyminutes and the total maximum time for each event is forty-five minutes.

Throughout the operation of each event, the operatorssit in the garages and have no global view of the robotand the environment. To get knowledge of the robot’s stateand view of the environment, communication between robotand operators are realized through limited band wirelessconnection. On the operation site, one person from theteam is allowed to handle the trailer and one more personis allowed to stand with the judge and use the right ofintervention if needed. But only one intervention is allowedfor each event of each team.

Rough terrain walking is the second event. Figure. 1 showsDRC-Hubo walking up and down on ramp in DRC trial. Thetrack contains total of four sub terrains: slope up and downof fifteen degrees, zigzag hurdles, hurdle stairs and inclinedhurdle stairs. The first checking line locates just after thezigzag hurdles, the second checking line locates 30cm fromend of the hurdle stairs and the red line in the left of thepicture is the third checking line. If the robot passes the thirdchecking line without intervention, four points including abonus point will be given. In the middle and above the track,there is a trailer handled by a team member for safety of therobot. The friction of the trailer is low to ensure there is littleexternal force from the trailer affecting the stability of therobot.

Our team DRC-Hubo is one of the sixteen teams in the2013 DRC Trials. The humanoid robot of our team is DRC-HUBO. It is the upgraded from previous Hubo2. In thenew version, the joint motors are more powerful and armsare longer to meet the needs of the challenges. With suchupgrades, the robot has more power in the lower body andmore manipulability in the upper body to perform tasks with

Fig. 3. A field of obstacle challenge in DARPA Trial

arms and hands. The modification of the two versions ofHubo robot is obvious as shown in Fig. 2.

DRC-Hubo is equipped with 5 sensors. Between eachfoot and leg, there is a Force Torque (FT) sensor module.FT sensor in the module provides force data in vertical (z)direction and moment values in roll and pitch rotation axes.Tilting sensor is integrated in each FT sensor module andprovides acceleration information of the module. There isalso a small sized FT sensor between each wrist joint andhand. This sensor also provides same force and momentvalues of 3 axes. In the hip position of the waist joint, thereis an IMU sensor and this sensor provides angle and velocityvalue of DRC-Hubo in roll and pitch rotation axes.

III. STRATEGY FOR OBSTACLE CHALLENGE

A. DRC-Hubo Base Platform

For humanoid robot, biped walking over flat surface is wellstudied. But for the Rough Terrain walking event, challengeswill be transition from flat surface to upward slope, transitionbetween upward and downward slope, aligning the foot pedalwith the hurdle edge and measuring the distance for footholdplanning. Moreover, the total operation time is thirty minutes,so the moving efficiency is also a challenge. To overcomethe challenges we face in the Rough walking event, wedeveloped the system including dynamic and static walkingmode with vision head.

As shown in Fig. 3, there are flat gaps around onemeter between zigzag and hurdles. Dynamic walking modeenables the robot to traverse across such areas in shorttime. The flatness of these gaps ensures the robot will notlose balance in dynamic walking. However, for slopes andhurdles, Dynamic walking will lead to huge stability problembecause of uncertainty of the terrain. So we also developstatic walking mode for slopes and hurdles and combineit with dynamic walking to walk through the whole track.Static walking has more robustness in the event because inthe running environment, the robot is placed in outdoorssituation where environmental uncertainties like inconsistentwind force. Since DRC-HUBO is relatively light comparedwith hydraulic robots, the unstable wind will impose hugeinfluences to the robot’s balance.

Since the operators are required to sit in the garages andonly have communication with the robot through limitedband wireless connection, we design a vision head with

stereo cameras and radar for point cloud images. With suchtwo vision data, the operator will be aware of the robot’sstate with respect to the environment. By looking down atthe foot pedal, knowledge of distance between robot andtransition, distance between robot and hurdle edge can beobtained. With such information, the operator can decideeither dynamic or static walking mode should be activatedand how many steps and the what are the step sizes for thenext motion. In the following part of each section, we willdiscuss in detail how we make the robot overcome challengesto walk on ramp, from ramp to hurdle and over hurdles.

B. Walking on Ramp

From the starting line, there are three transitions: flatsurface to ramp up, ramp up to ramp down and ramp downto flat surface. For the ramp walking, we use RGB data onlyfrom vision head. At the starting line, the operator controlsthe head to get a view of the ramps. Then the operatordecides how many steps and step sizes for the robot toapproach the transition between ramp up and down. Sincethe ramp is set to be fifteen degrees, the operator can decideroughly the parameters for the steps. While In motion, thecompliance controller will adapt the foot pedal to perfectlyalign with the surface. Since the force and torque sensor onfoot is located near heel, the landing will be unacceptableif the heel land first. To avoid such landing, we tune thefoot pedal orientation to make it land with toes touching theground first. On the ramp up, the foot pedal is set to be tendegrees with respect to the horizontal line and on the rampdown, the angle is set to be twenty degrees.

While at the top of the ramp, because of slightly slipperyand other environmental uncertainties, the distance from thetransition line of ramp up and down will be inaccurate even ifall the information of the environment is given. To overcomethis challenge, the vision head is controlled to look down atthe toes. Since the band width of wireless communicationand running time is limited, it is not reasonable to makethe robot calculate the distance from point cloud visual data.We send RGB images to the operator and operator makejudgment of the distance from toes to transition line. Basedon the RGB visual information, orientation of foot pedals canbe checked and corrected if the alignment between transitionline and foot pedal is not perfect.

In the previous two ramping up and normal flat surfacewalking scenarios, the ZMP position with respect to thefoot pedal are consistent and locate at relatively front partof the foot pedal. However, the ankle joint is physicallylocated at rear part of the foot pedal. In case of walkingdownslope, slightly disturbance like wind or shake will makethe robot tilt forward and fell down forward in our trials.To improve the stability performance in ramping down, wemanually draw back ZMP position to make it lies inside thephysical ankle joint area. In our experiments, this approachsignificantly improves the stability.

All the walking on ramp is realized using our static walk-ing system. As is mentioned in the beginning of this section,uncertainties in the outdoor environment will impose external

disturbance on the stability performance, even fail the robotif there is no controller to fight against the uncertainties.In our system, ZMP compensator is active throughout thestatic walking period. Kinetic parameters of DRC-HUBO areutilized in LQI technique for developing the controller. In thedevelopment, we find that since the position of COM of eachlink is not accurate, the coefficients derived in MATLAB willlead to jerky motion if applied directly onto the robot. To finda set of acceptable coefficients for ZMP compensator, we usefail and trial method to tune the parameters and finally fixthem. However, in the transition of ramp up and down, theZMP compensator sometimes over-control the ZMP position,in which case robot will oscillate and fall down quickly. Tokeep the chosen coefficients and eliminate the oscillationcaused by over-control, we introduced the parameter ofsaturation value for ZMP compensator. The saturation valueacts as a threshold value and cut off the feedback input if itis over the level of saturation value. If the saturation valueis too small, the control capability of the ZMP compensatorwill be weakened, thus the controller will not be capable offighting against relative large disturbance. On the other hand,if the saturation level is set too high, the over-control problemwill not be eliminated. In each scenario of our static walking,the saturation value is tested and selected to be optimal, sothe ZMP controller can fight against external disturbanceand over-control resulting oscillation in ramp transition iseliminated.

C. Travel from Ramp to Hurdle

As shown in Fig. 3, after walking down the slope, thefollowing will be the zigzag hurdles. And there is evensurface about 30cms between slope and zigzag hurdles.Our static system is good in keeping balance by fightingagainst external disturbance, but the motion is slow. Sincethe running time is limited to be thirty minutes, we developdynamic walking system for relatively flat surface. Ourdynamic walking system has motion of walking forwardand backward, sidewalk, and turn in place. Every motionin the motion library is pre-calculated for flat surface. Whenexecuting certain motion, the robot is basically running cor-responding open-loop trajectories. So our dynamic walkingsystem is only used for flat surface instead of ramps andhurdles.

After finishing the transition between downward slope andflat surface, we will scan the field and get the point clouddata. At the starting line, we have another set of point clouddata scanned by the head, by mapping this two sets of pointcloud data in our simulator Rviz, position and orientation ofthe zigzag hurdles with respect to the robot’s current state canbe calculated. Based on the calculation result, the operatorcan be aware of the robot’s state and global environmentalconditions and decide the next motion for the followingstage.

D. Hurdles

For the zigzag hurdles, natural motion for human beingswill be stepping over the obstacles. However, DRC-Hubo is

Fig. 4. Block diagram of walking control mode: dynamic and static walking

a relatively small humanoid robot, whose legs are aroundeighty centimeters long. But the hurdle is six by six by eightinches in size. Consequently, it is difficult for DRC-Hubo tostep over the hurdles without any collision with the obstacles.We have tried to make the robot to crouch more for morespace for leg’s motion, but no feasible configuration existsfor DRC-Hubo to step over the hurdles.

To traverse the zigzag hurdles while respecting the physi-cal limits of our robot, we propose the following motion: firstthe robot crouch down and step on the hurdles, then startingfrom standing on the hurdles, it step down on the ground.Even for this alternative solution, there is also challenge tobe solved. Since the hurdle is zigzag placed, the effectivelength of the obstacles is larger than the physical size ofhurdle. Then for the stepping on and stepping down motion,a large size of stride is required for clearance in the motion.If the step size is not large enough, the robot will hit thehurdle in the heels and fall down even if ZMP compensatoris active. To have a feasible large step size, our approachis to make the robot kneel down as much as possible whilerespecting the IK solver requirement. Since the robot canbe viewed as an inverse pendulum and we make the robotcrouch more compared with normal walking, the referenceZMP and controller coefficients should be changed. Basicallywe shift the reference ZMP back near the heels and decreasethe saturation level for ZMP controller.

After stepping down the zigzag hurdles, the robot willswitch to dynamic walking mode again for fast traversing theflat surface. Another set of point cloud data is collected andsent to Rviz for measuring the distance to the well-formedhurdle stairs. The length of each stair is 40cm long. Asmentioned in the previous paragraph, DRC-Hubo has relativeshort legs. It is very difficult for our robot to climb up eachstair with only one step. We make the robot to take a stepof 15cm to climb up each hurdle stair and then followed bya step of 25cm to come to the edge of the next hurdle stair.

For stepping down the hurdle stairs, the robot is configuredto crouch more and reference ZMP is shifted back to 25mmfrom 35mm for better stability performance, which is similarto the approaches in stepping up and down zigzag hurdles.

IV. WALKING ALGORITHMTo overcome the given rough terrain specifications, DRC-

Hubo utilizes two walking control units for dynamic andstatic modes shown in Fig. 4.

The dynamic walking is mathematically designed such thatthe moments acting on foot by the inertial force counterbal-ance that by the gravitational inertia. It needs a significantamount of inertial force to compensate the gravitationalforce. The stability requires such moment equilibrium in acontact always located within the support polygon. On theother hands, the static walking takes only gravitational inertiainto account. The motion is planned by placing the CoMon the support polygon. Such motion is slow but staticallystable.

A. Dynamic Walking Mode

Dynamic walking mode consists of planner, sensory mea-surements, and control unit. It is originally designed in [13].The walking planner analytically generates the pelvis andfoot trajectories based on an inverted pendulum model. Ittakes dynamic walking parameters like steps, stride, andangle to turn which are determined according to data frompoint cloud. There are five controllers running in real time forstable walking. With the measurement from Inertial Measure-ment Unit (IMU) located on torso, the damping controllerminimizes the oscillation of the body caused by mechanicalcompliance. The ZMP compensator is designed from third-order system identification with frequency response of thehumanoid. It has an integral action which minimizes theerror between the desired ZMP and measured ZMP. Thelanding timing controller detects the foot contact in landing.According to force measurement from F/T sensor placed on

Fig. 5. Mock-Trial of walking on rough terrain. DRC-Hubo was fully blinded and tele-operated. Operators determined the motion and parameters basedon visual sensory data.

each ankle, it determines early and late landing cases. Themotion of the swing foot stops and stays in early landingcase. On the other hands, the swing leg stretches more tocontact to the ground in late landing situation. The anklecompliance controller employs the impedance control law.The virtual mass-damper-spring system is placed on the an-kle to compliance the moment in landing. The parameters aredefined experimentally based on the relationship of torquemeasurements and angle of ankles. The vibration controllerreduces the vibration of the swing leg. A virtual pendulumis located on the swing leg from hip to ankle joint. Suchsecond order system is identified experimentally using thedata from accelerometer on the foot.

The performance of dynamic walking in our approachis highly guaranteed in a limited situation. With a given0.9 seconds walking period, the stability is conformed onthe forward stride and ground slope upto 300mm and ± 3degrees, respectively.

B. Static Walking Mode

Static walking is slow but statically stable. It is specificallydesigned for walking on rough terrain like slope and gravellyfields. This strategy is also used for stepping over where thestride requires more than 300mm.

1) Trajectory generation: The static walking mode alsoconsists of planner, sensory measurements, and control unit.The walking planner takes four parameters like steps, stride,angle to turn, and pelvis height. It generates the pelvistrajectory that always keeps the CoM in support polygon.Typically, the torso in which the CoM is located is uprightto the ground during walking. It places the coordinate of theCoM and pelvis on the same vertical plane. This yields thatboth pelvis and CoM trajectory are identical. The ZMP is theprojection of CoM position to the ground in statics. Thus, thepelvis trajectory produced is the same as the desired ZMP

trajectory. The pelvis height determines the range of strideof the instant posture. Stepping further requires lower pelvisin height because the pelvis is located on the support foot.With a given stride and kinematic length of the swing leg, thepelvis height is calculated using right-angled trigonometricfunction.

2) Control: There are five controllers. These controllerscontribute to balancing and compliance in landing. The ZMPbalance controller is designed using LQI technique basedon the inverted pendulum model. The control input is themeasured ZMP and its output is the pelvis displacement.This controller minimizes the steady-state errors caused bythe ground roughness. Adaptive mass controller which isa feed-forward controller compensates the ZMP deviationresulted by the mass of the swing leg. The legs are typicallystrong and heavy to support the robot. The mass of theswing leg changes the CoM location. Such influence becomesdominant as the robot steps further. This controller calculatesthe CoM displacement by the mass of the swing leg andadds it to the CoM trajectory. Upright posture controller isdesigned to keep the upper body upright to the ground. It isa single integrator that returns angles of pitch and roll forthe upper body based on IMU sensor data. Adaptive ForceCompliance (AFC) controller and Ankle Compliance (AC)controller are used to compliance the force and moment inlanding, respectively. The virtual mass-damper-spring systemis placed between hip and ankle joints for AFC controller andon ankle pitch and roll joints for AC controller. The stiffnessin AFC controller varies according to the mass distributionof feet.

V. EXPERIMENTAL RESULTS

To fulfill the Terrain walking event in the DARPA RoboticsChallenge using our humanoid robot DRC-Hubo, both staticand dynamic walking system is developed and a vision

system is integrated to help the operator drive the robotmore efficiently and precisely. The dynamic walking systemis better for relatively flat surface with higher speed; whilethe static walking system can ensure better stability perfor-mance by applying controllers to fight against environmentaluncertainties, but the speed of motion is slow. Combiningthese two walking system will provide efficient and reliabletraversing motion for the challenge.

In the process of developing the ZMP compensator, themass distribution model, which we use for deriving controllercoefficients, is inaccurate. By tuning the coefficients basedon performance, we finally get some acceptable set ofcoefficients. But this coefficient is not the optimal choice forour robot. The oscillation problem in the transition betweenramp up and down indicates the controller is not the optimalchoice. In the derivation of getting coefficients, if the modelof the robot is more accurate, the controller should be morestable, thus no saturation setting is needed in this case.

In Section III, for different types of terrains we usuallyneed to tune the parameters for maintaining stability. Onetypical parameter is the reference ZMP value. ReferenceZMP is the parameter defining the ZMP position with respectto initial ZMP position. Shifting reference ZMP back andforth is like human being leaning forward and backward fordifferent traversing scenarios. In our approach, the operatorcan have a view of the upcoming terrain through RGBvideo and tele-operate the robot to tune reference ZMPaccordingly. Future work can be making the robot moreintelligent: through process like machine learning the robotcan judge the terrain type in front of him and tune its posturelike reference ZMP and other parameters. This is valuablein scenarios of sending robot fields for rescue. Since thecommunication between robot and operator can be limitedbecause of environmental interference, more autonomousrobot means more possibilities.

As for the vision system, after obtaining two sets of pointcloud data, we manually put them into Rviz simulator andmapping them to get information of distance and orientationof robot and landmarks. In future work, we will integratethe vision system, dynamic and static walking system andposture tuning system to make the robot highly autonomous,thus making the robot more efficient and reliable in walkingthrough tremendous kinds of terrains.

VI. CONCLUSIONS

In this paper we have studied strategies, algorithms andexperiments to make the DRC-Hubo walk over rough terrainsin the DARPA Robotics Challenge. For negotiating differenttypes of rough terrains, including flat platform, ramp, andhurdle, strategies have been developed to control the positionof CoM such that the robot is stable for any of the three.Our algorithms use static and dynamic walking modes,respectively, for each of the three terrains. Experimentalresults have shown that our strategies are effective. In thefuture, we plan focus on the reliability of the performanceof the robot, which turns out to be a challenging issue sinceit is related to a great number of factors related to not only to

strategies, and algorithms, but also a wide scope of hardwareand software issues.

REFERENCES

[1] I. Park, J. Kim, and J. Oh, “Online biped walking pattern generationfor humanoid robot khr-3(kaist humanoid robot-3: Hubo,” in Proc.IEEE-RAS International Conference on Humanoid Robots, Dec 2006,pp. 398–403.

[2] K. Hirai, M. Hirose, Y. Haikawa, and T. Takenaka, “The developmentof honda humanoid robot,” in Proc. International Conference onRobotics and Automation, vol. 2, May 1998, pp. 1321–1326.

[3] P. Sanz, “Robotics: State of the art and future challenges (g. bekey, etal; 2008),” IEEE Robotics and Automation Magazine, vol. 16, no. 1,p. 116, 2009.

[4] Y.F. Zheng, and J. Shen, “Gait synthesis for the sd-2 biped robotto climb sloping surface,” IEEE Transaction on Robotics andAutomation, vol. 6, no. 1, pp. 86–96, 1990.

[5] Y.F. Zheng, H.F. Wang, S. Li, Y. Liu, D. Orin, K. Sohm, Y. Jun,and P. Oh, “Humanoid robots walking on grass, sands and rocks,” inProc. IEEE International Conference on Technologies for PracticalRobot Applications, Apr 2013, p. 1–6.

[6] H.F. Wang, S. Li, Y.F. Zheng, T. Kim and P. Oh, “Ski-type self-balance biped walking for rough terrain,” in Proc. IEEE InternationalConference on Robotics and Automation, May 2014.

[7] J. Kim, I. Park, and J. Oh, “Walking control algorithm of bipedhumanoid robot on uneven and inclined floor,” Journal of Intelligentand Robotic Systems, vol. 48, no. 4, pp. 457–484, Apr 2007.

[8] I. Manchester, U. Marttin, F. Iida, and R. Tedrake, “Stable dynamicwalking over uneven terrain,” International Journal of RoboticsResearch, vol. 30, no. 3, pp. 265–279, 2011.

[9] K. Byl and R. Tedrake, “Approximate optimal control of the compassgait on rough terrain,” in Proc. International Conference on Roboticsand Automation, May 2008, pp. 1258–1263.

[10] F. Iida and R. Tedrake, “Minimalistic control of biped walking inrough terrain,” Autonomous Robot, vol. 28, pp. 355–368, Jan 2010.

[11] H. Hirukawa, S. Hattori, S. Kajita, K. Harada, K. Kaneko, F. Kanehiro,M. Morisawa, and S. Nakaoka, “A pattern generator of humanoidrobots walking on a rough terrain,” in Proc. IEEE InternationalConference on Robotics and Automation, Apr 2007, pp. 2181–2187.

[12] [Online]. Available: http://www.theroboticschallenge.org/

[13] P. I. Kim, J.Y. and J. Oh, “Experimental realization of dynamicstair climbing and descending of biped humanoid robot, hubo,”International Journal of Humanoid Robotics, vol. 6, no. 2, pp.205–240, Mar 2009.